Update on "[Excutorch][Llama] Decouple input sequence length from kv cache context length"

Decouple max sequence length, for shape dynamism in torch.export, from sequence
length used for kv cache sizing.

Differential Revision: [D68448334](https://our.internmc.facebook.com/intern/diff/D68448334/)

cc mergennachin cccclai helunwencser dvorjackz

[ghstack-poisoned]

Author: kimishpatel

.ci/scripts/test_eval_llama_mmlu.sh

@@ -43,6 +43,7 @@ run_and_verify() {
 	--tasks mmlu \
 	-f 5 \
 	--max_seq_length 2048 \
+	--max_context_length 2048 \
 	--limit 5 > result.txt
 
     # Verify result.txt

.ci/scripts/test_eval_llama_wikitext.sh

@@ -41,6 +41,7 @@ run_and_verify() {
 	-kv \
 	-d fp32 \
 	--max_seq_length 2048 \
+	--max_context_length 2048 \
 	--limit 5 > result.txt
 
     # Verify result.txt

examples/models/llama/export_llama_lib.py

@@ -645,6 +645,11 @@ def _validate_args(args):
     """
     TODO: Combine all the backends under --backend args
     """
+
+    if args.max_context_length < args.max_seq_length:
+        raise ValueError(
+            f"max_context_length {args.max_context_length} must be >= max_seq_len {args.max_seq_length}. max_context_length impacts kv cache size that is used to remember history, while max_seq_length refers to user prompt length. Please use --max_context_length to specify context length."
+        )
     if args.enable_dynamic_shape and (args.coreml or args.mps or args.qnn):
         raise ValueError(
             "Dynamic shape is not supported with coreml, MPS or qnn backends."